PydanticAI

Pydantic AI is a Python library (pip install pydantic-ai) for building LLM agents with typed structured outputs, typed dependency injection, and typed tool calling — the same design philosophy as the Pydantic validation library, which is used inside the OpenAI SDK, Anthropic SDK, Google ADK, LangChain, and FastAPI. Models are swappable with a string ('anthropic:claude-fable-5', 'openai:gpt-5.6-sol', etc.) across OpenAI, Anthropic, Google, Bedrock, Azure AI Foundry, Groq, Mistral, xAI, Ollama and others, with no flagship feature locked to one vendor, per the vendor's own docs.

The typing is real, not decorative: structured outputs, dependency injection, and tool signatures are checked by mypy/pyright, which is a concrete advantage over prompt-templated or dict-based agent frameworks for teams that already have Python type discipline. It has one of the more credible 'not locked to one vendor' stories in this space — model swap is a string change, and there's a built-in offline 'test' model for exercising agent logic without hitting an API or paying for tokens. First-party, co-maintained durable-execution integration with Temporal, DBOS, and Prefect (plus community integrations for Restate and Airflow) is unusual — most agent frameworks leave durability entirely to the user.

19,399 GitHub stars and 2,561 forks as of 2026-08-20, created June 2024. GitHub's dependency graph shows 4,492 public repositories and 894 packages depending on it. Release cadence is genuinely daily-to-every-few-days (v2.32.0, v2.31.1, v2.31.0 shipped on three consecutive days, Aug 17–19 2026), and PRs are merging the same day this was checked — this is not a project coasting on stars. Thoughtworks' Technology Radar (Nov 2025) placed it in the 'Trial' ring, describing it as a 'stable, well-supported, open-source framework for building GenAI agents in production' — the one piece of independent analyst coverage found, as opposed to vendor-published case studies.

Four GitHub Security Advisories have been filed against pydantic-ai in 2026, all the same vulnerability class (CWE-918, Server-Side Request Forgery) in URL-download handling used by web-facing integrations (Agent.to_web, VercelAIAdapter, AGUIAdapter): GHSA-2jrp-274c-jhv3 (HIGH, published Feb 2026, fixed 1.56.0), GHSA-wjp5-868j-wqv7 (HIGH, path traversal/XSS, Feb 2026, fixed 1.51.0), GHSA-cqp8-fcvh-x7r3 (MODERATE, May 2026, fixed 1.99.0), GHSA-cg7w-rg45-pc59 (MODERATE, Jun 2026, fixed 1.102.0), and GHSA-h7p7-w5gc-xj3w (MODERATE, published Aug 13 2026, fixed 1.106.0 / current release line). Each was responsibly disclosed and patched, generally within the same or next minor release. Only applications that accept message history or URLs from untrusted users are affected — a backend agent with hardcoded, developer-controlled URLs is not.

The core library and the newer pydantic-ai-harness package (v0.23.0, also releasing multiple times a week) are both MIT-licensed with no paid tier gating any framework feature — the vendor states directly that 'the Pydantic data validation library is (always has been, always will be) completely free and permissively licensed under the MIT license,' and that applies to Pydantic AI too. The only paid product is Pydantic Logfire, an optional OpenTelemetry-based observability platform from the same company: free Personal tier (10M records/month, hard-capped, no card required), Team at $49/mo (5 seats included, $25/extra seat, $2/million additional records), Growth at $249/mo (unlimited seats, up to 90-day retention, HIPAA BAA), and custom Enterprise (cloud, dedicated, or self-hosted, SOC 2 Type II, SSO). Logfire's built-in AI Gateway charges a markup (5% on Personal/Team, 3% on Growth, 0% if you bring your own provider keys) if you route model calls through it rather than calling providers directly — none of this is required to use the pydantic-ai library itself.

Pydantic Services Inc. is a venture-backed company founded by Samuel Colvin, who created Pydantic in 2017. It raised a $4.7M seed round led by Sequoia Capital in Feb 2023, then a $12.5M Series A led by Sequoia (with Partech, Irregular Expressions, and angels including Logan Kilpatrick and Jason Liu) in Oct 2024, alongside the Logfire GA launch — roughly $17.2M raised total, no valuation disclosed in either announcement. TechCrunch reported 13 employees across the US and Europe as of the Series A; later headcount figures from third-party trackers (Tracxn, PitchBook) are self-reported estimates, not independently confirmed, so treat any number above 13 as unverified.